-
Notifications
You must be signed in to change notification settings - Fork 85
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
create Argilla
integration for haystack
nodes
#25
create Argilla
integration for haystack
nodes
#25
Comments
|
@TuanaCelik, what do you think about this issue? |
@davidberenstein1957 this to me sounds like an amazing idea. @masci @anakin87 wdyt? In terms of usage, I think the most interesting one might be |
It seems great to me! I haven't tried Argilla Feedback and Callbacks but I think that by setting up an ArgillaCallbackHandler,
|
@anakin87 would you see that as it's own component, or would you see it as a functionality of a component. |
I am interested in working on this. I am not sure if a separate package makes sense or we should install it through extras? But let me know what makes sense for your part of the codebase. |
Hey @davidberenstein1957 - we're going to do some investigation with @anakin87 on Monday to get a better understanding how/if the source code should change, or whether this would be possible to implement as a lone standing integration. We will comment back here with our findings 👍 |
No problem👍 |
Hey @davidberenstein1957 we've done some thinking and we see 2 options for you. And it's largely your choice of what type of integration you would like to build:
If you decide to go with option 2, @anakin87 had the idea that it could even be a custom node that can wrap another node like the We're currently desigining Haystack v2, whatever option you prefer to go for, we would be happy to update you if there's any major change that would make your integration better. E.g., if we do end up implementing callbacks for PromptNode etc. |
Hi @TuanaCelik @anakin87 , I think starting off with the callback functionality would be best in terms of effort vs gain. However, I think an ArgillaNode encompassing a lot of additional tasks would be a more robust solution for the future but we work on this, whenever the first integration proves successful. What do you think? |
Thanks, @davidberenstein1957! |
@anakin87 Great! So, from your perspective, should we add the |
Sorry! I missed this one... |
@anakin87 - correct me if I'm wrong, but I think currently this would be a good integration and if @davidberenstein1957 is interested in it, we can involve you in Haystack v2 discussions when we start looking into evaluation and observability. You can see our current status on the Discussions page. |
@TuanaCelik @anakin87, I noticed that you do not directly allow for customization of the For me, it would make sense to change the And introduces a base CallbackManager class and allow for passing a custom implement of this class to the from abc import ABC, abstractmethod
class BaseCallbackManager(ABC):
@abstractmethod
def abstract_func(self, *args, **kwargs):
pass
class CallbackManager(BaseCallbackManager):
def __init__(self, agent_color: Color = Color.GREEN) -> None:
self.agent_color = agent_color
def on_tool_finish(
self,
tool_output: str,
color: Optional[Color] = None,
observation_prefix: Optional[str] = None,
llm_prefix: Optional[str] = None,
**kwargs: Any,
) -> None:
print_text(observation_prefix) # type: ignore
print_text(tool_output, color=color)
print_text(f"\n{llm_prefix}")
def on_new_token(self, token: str, **kwargs: Any) -> None:
print_text(token, color=self.agent_color)
def on_agent_start(self, **kwargs: Any) -> None:
agent_name = kwargs.pop("name", "react")
print_text(f"\nAgent {agent_name} started with {kwargs}\n", color=self.agent_color)
def on_agent_step(self, agent_step, **kwargs: Any) -> None:
print_text(agent_step.prompt_node_response, end="\n", color=self.agent_color) I, and others, would then be able to do: class ArgillaCallbackManager(CallbackManager):
def __init__(self, *args, **kwargs):
pass
argilla_callbackmanager = ArgillaCallbackManager()
prompt_node = ...
agent = Agent(
prompt_node=...,
callback_manager=argilla_callbackmanager
) Let me know if I can pick this up in a PR to the core package. |
I also noticed that |
@davidberenstein1957 your observations make sense to me and I would say that introducing a base |
For me, it makes sense to also add a couple of additional callback points and use the same callback manager for both agents and tools from the |
@davidberenstein1957 apologies for the late response; I never got these notifications you sent. I understand your intentions regarding the custom Argilla callback manager. In striving for simplicity and prevention of a class explosion, we used the Events class from events Python library for agent callbacks. Is there a way you can adapt to this approach for the 1.x branch, and then for 2.x, we can talk more about trade-offs of a more explicit callback mechanism and which one to use? Let us know your thoughts. |
@vblagoje I think this normally would be a problem but the hardcoded event-call are still limited. At the moment it is not possible to get the final answer from the Events because there is no In order to do the I propose altering this line https://github.com/deepset-ai/haystack/blob/d048bb53523b11426d3c39f7cdc0aa8d15f0cf67/haystack/agents/base.py#L362 |
That seems reasonable @davidberenstein1957 Would you please follow up on this idea @anakin87 ? |
@anakin87 I would like to handle the PR. Do you think it makes sense to also include other steps outlined by |
@davidberenstein1957 As this part will be rethought and redesigned in Haystack 2.0, I would say open a PR and include the minimum set of events needed for the Argilla implementation. |
@anakin87 I outlined the PR. Not sure if it requires doc changes or tests? |
@anakin87 @TuanaCelik , do you have any pointers w.r.t. this PR? We would love to log some more info into Argilla but maybe the Agent implementation is still limited? We hope to align further with |
Hi, @anakin87 @TuanaCelik, finally we decided to add the integration for Haystack 1.x with Argilla. But we're excited about the Haystack 2.0 possibilities and we plan a refactoring when agents or callbacks become available (I think that it's planned for the upcoming months). So, we are looking forward to the new updates! :) |
I think it would be cool to be able to monitor Haystack pipelines and nodes directly to Argilla.
https://docs.haystack.deepset.ai/docs/prompts_and_llms could be similar to https://docs.argilla.io/en/latest/guides/llms/practical_guides/use_argilla_callback_in_langchain.html
However, I can also see support for:
The text was updated successfully, but these errors were encountered: